AAAI 2026 Papers

Ultra-high-resolution (UHR) text-to-image synthesis faces significant hurdles, including immense computational costs and a scarcity of training data. To address these, we introduce RealUHR, an efficient and scalable framework for generating photorealistic 4K images. At its core, RealUHR employs a Patch-Cascade Flow Matching pipeline that ensures global coherence without costly patch fusion by initiating generation from a semantically meaningful structure. This enables highly efficient, few-step inference for independent patches. Our key contribution is Guidance-Consistent Adaptation (GCA), a novel two-stage strategy to resolve the fundamental objective mismatch in guidance-distilled models. GCA allows powerful backbones like FLUX to be effectively adapted for patch-aware UHR synthesis. The framework's detail-rendering capabilities are further enhanced by a non-uniform time schedule. Experiments show that RealUHR establishes superior performance in both quality and efficiency, and excels in zero-shot applications such as creative up-sampling and generative artifact suppression.

Existing multi-person video pose estimation methods typically adopt a two-stage pipeline: detecting individuals in each frame, followed by temporal modeling for single-person pose estimation. This design relies on heuristic operations such as tracking, RoI cropping, and non-maximum suppression, limiting both accuracy and efficiency. In this paper, we present a fully end-to-end framework for multi-person 2D pose estimation in videos, effectively eliminating heuristic operations. A key challenge is to associate individuals across frames under complex and overlapping temporal trajectories. To address this, we introduce a novel Pose-Aware Video transformEr Network (PAVE-Net), which features a spatial encoder to model intra-frame relations and a spatiotemporal pose decoder to capture global dependencies across frames. To achieve accurate temporal association, we propose a pose-aware attention mechanism that enables each pose query to selectively aggregate features corresponding to the same individual across consecutive frames. Additionally, we explicitly model spatiotemporal dependencies among pose keypoints to improve accuracy. Notably, our approach is the first end-to-end method for multi-frame 2D human pose estimation. Extensive experiments show that PAVE-Net substantially outperforms prior image-based end-to-end methods, achieving a 6.0 mAP improvement on PoseTrack2017, and delivers accuracy competitive with state-of-the-art two-stage video-based approaches, while offering significant gains in efficiency.

How far are deep models from real-world video anomaly understanding (VAU)? Current works typically emphasize detecting unexpected occurrences deviating from normal patterns or comprehending anomalous events with interpretable descriptions. However, they exhibit only a superficial comprehension of real-world anomalies, with limited breadth in complex principles and subtle contexts that distinguish the anomalies from normalities, e.g., climbing cliffs with safety gear vs. without it. To this end, we introduce CueBench, the first of its kind Benchmark, devoted to Context-aware video anomalies within a Unified Evaluation framework. We comprehensively establish an event-centric hierarchical taxonomy that anchors two core event types: 14 conditional and 18 absolute anomaly events, defined by their refined semantics from diverse contexts across 174 scenes and 198 attributes. Based on this, we propose to unify and benchmark context-aware VAU with various challenging tasks across recognition, temporal grounding, detection, and anticipation. It also serves as a rigorous and fair probing evaluation suite for generalized and specialized vision-language models (VLMs) across both generative and discriminative paradigms. To address the challenges underlying CueBench, we further develop Cue-R1 based on R1-style reinforcement fine-tuning with verifiable, task-aligned, and hierarchy-refined rewards in a unified generative manner. Extensive results on CueBench reveal that, existing VLMs are still far from satisfactory real-world anomaly understanding, while our Cue-R1 surpasses these state-of-the-art approaches by over 24% on average.

Recent advances in image editing tools, particularly those used in content-aware retouching and object-level manipulation, have raised significant concerns regarding the authenticity of digital images. While many Image Manipulation Detection and Localization (IMDL) methods have been proposed, they often struggle with subtle forgeries, intricate boundary artifacts, and manipulations generated by unseen editing techniques. In this work, we propose a novel edge-aware framework that leverages the strong natural image priors of pre-trained inpainting models to harmonize manipulated regions. By guiding the inpainting process with generated edge-aware masks, our method reconstructs tampered areas using surrounding context, yielding perceptually coherent results. The pixel-wise residual between the original and reconstructed images reveals manipulation-sensitive inconsistencies—particularly around editing boundaries—thereby enabling accurate and generalizable detection and localization. Extensive experiments across multiple benchmarks demonstrate that our approach achieves state-of-the-art performance, especially in challenging scenarios involving realistic and finely retouched image forgeries.

Point cloud tasks have recently benefited from Mamba-based architecture, which leverage state space modeling to achieve strong performance. Previous studies have primarily focused on network design while overlooking the importance of position encoding and relying on coarse-grained geometric feature aggregation. The former leads to semantic ambiguity due to inconsistent spatial relationships, while the latter results in geometric feature dispersion by overlooking fine-grained local geometric details. To tackle the above problem, we propose a novel framework, PointMC, including Multi-view Consistent Learnable Position Encoding (MCLPE) and Center-Global Feature Fusion (CGFF), to provide semantically coherent positional guidance for inter-patch and enable fine-grained geometric structure aggregation within intra-patch regions. Specifically, the proposed MCLPE module is inspired by a spatial structure modeling mechanism guided by physical constraints, leverages multi-view virtual reconstruction and a learnable strategy to dynamically constrain spatial relationships along patch boundaries, thereby enhancing the semantic consistency and representational clarity across inter-patch regions. Furthermore, considering the lack of local structural information within each patch, the CGFF module employs a dual-guidance mechanism based on center and global structures to effectively promote the aggregation of local geometric features. Extensive experiments on multiple benchmark datasets validate the effectiveness of PointMC, consistently outperforming existing state-of-the-art methods, and demonstrating superior capability in capturing both inter-patch semantic consistency and intra-patch geometric details.

Lung cancer continues to be the leading cause of cancer-related deaths globally. Early detection and diagnosis of pulmonary nodules are essential for improving patient survival rates. Although previous research has integrated multimodal and multi-temporal information, outperforming single modality and single time point, the fusion methods are limited to inefficient vector concatenation and simple mutual attention, highlighting the need for more effective multimodal information fusion. To address these challenges, we introduce a Dual-Graph Spatiotemporal Attention Network, which leverages temporal variations and multimodal data to enhance the accuracy of predictions. Our methodology involves developing a Global-Local Feature Encoder to better capture the local, global, and fused characteristics of pulmonary nodules. Additionally, a Dual-Graph Construction method organizes multimodal features into inter-modal and intra-modal graphs. Furthermore, a Hierarchical Cross-Modal Graph Fusion Module is introduced to refine feature integration. We also compiled a novel multimodal dataset named the NLST-cmst dataset as a comprehensive source of support for related research. Our extensive experiments, conducted on both the NLST-cmst and curated CSTL-derived datasets, demonstrate that our DGSAN significantly outperforms state-of-the-art methods in classifying pulmonary nodules with exceptional computational efficiency.

Current novel view synthesis methods are typically designed for high-quality and clean input images. However, in foggy scenes, scattering and attenuation can significantly degrade the quality of rendering. Although NeRF-based dehazing approaches have been developed, their reliance on deep fully connected neural networks and per-ray sampling strategies leads to high computational costs. Furthermore, NeRF's implicit representation limits its ability to recover fine-grained details from hazy scenes. To overcome these limitations, we propose DehazeGS, the first physics-driven 3D Gaussian Splatting (3DGS) framework for dehazing. We adopt an explicit Gaussian representation to model fog formation via a physically consistent forward rendering process, enabling reconstruction and rendering of fog-free scenes using only multi-view foggy images as input. Specifically, based on the atmospheric scattering model, we simulate the formation of fog by establishing the transmission function directly on Gaussian primitives via depth-to-transmission mapping. During training, we jointly learn the atmospheric light and scattering coefficients while optimizing the Gaussian representation of foggy scenes. At inference time, we remove the effects of scattering and attenuation in Gaussian distributions and directly render the scene to obtain dehazed views. Experiments on both real-world and synthetic foggy datasets demonstrate that DehazeGS achieves state-of-the-art performance.

Camera-based multi-view 3D detection is crucial for autonomous driving. PETR and its variants (PETRs) excel in benchmarks but face deployment challenges due to high computational cost and memory footprint. Quantization is an effective technique for compressing deep neural networks by reducing the bit width of weights and activations. However, directly applying existing quantization methods to PETRs leads to severe accuracy degradation. This issue primarily arises from two key challenges: (1) significant magnitude disparity between multi-modal features—specifically, image features and camera-ray positional embeddings (PE), and (2) the inefficiency and approximation error of quantizing non-linear operators, which commonly rely on hardware-unfriendly computations. In this paper, we propose FQ-PETR, a fully quantized framework for PETRs, featuring three key innovations: (1) Quantization-Friendly LiDAR-ray Position Embedding (QFPE): Replacing multi-point sampling with LiDAR-prior-guided single-point sampling and anchor-based embedding eliminates problematic non-linearities (e.g., inverse-sigmoid) and aligns PE scale with image features, preserving accuracy. (2) Dual-Lookup Table (DULUT): This algorithm approximates complex non-linear functions using two cascaded linear LUTs, achieving high fidelity with minimal entries and no specialized hardware. (3) Quantization After Numerical Stabilization (QANS): Performing quantization after softmax numerical stabilization mitigates attention distortion from large inputs. On PETRs (e.g., PETR, StreamPETR, PETRv2, MV2d), FQ-PETR under W8A8 achieves near-floating-point accuracy (

Diffusion models have advanced fine-grained garment generation, yet balancing controllability, efficiency, and texture fidelity remains challenging. Adapter-based methods often yield incoherent details, while full fine-tuning is computationally expensive and prone to overwriting pretrained priors. To address these limitations, we propose IMAGGarment+, an efficient diffusion framework for controllable and high-quality garment synthesis. It comprises two key modules designed for efficient and attribute-aware conditioning. First, we introduce an attribute-wise feature extractor (AFE) that disentangles key garment attributes, silhouette, logo, position, and color, into parallel latent streams. Each stream is optimized independently via LoRA, ensuring minimal parameter overhead while retaining expressive capacity. Second, we develop an attribute-adaptive attention (AA) module to inject attribute-specific cues into the generative process through a selective, layer-wise injection strategy. Specifically, silhouette and color features are injected into early decoder layers to guide structural and appearance formation, while logo features are propagated across all layers to ensure cross-scale consistency. Extensive experiments on fine-grained garment benchmarks demonstrate that IMAGGarment+ outperforms state-of-the-art baselines with less than 20% additional parameters, validating its effectiveness and efficiency.

Prediction of pedestrian behavior is crucial for autonomous driving systems and intelligent transportation.Conventional methods predict the behavior based solely on either the pedestrian intention or the distance-related interactions between the pedestrian and its surroundings. However, these methods overlook the associations between intention and interaction for behavior prediction, in which they should be aligned with each other, thus leading to sub-optimal predictions. To solve this problem, we propose to predict the behavior by learning the association between intention and interaction, enabling them to mutually enhance each other during the prediction. Specifically, we first predict the short-term intention of all objects, including the target pedestrian and its surroundings.Then, instead of using the distance-related interactions, we predict the interactions by learning the correlated intentions. Finally, the intention-driven interactions refine the initial intention prediction, thus ensuring the alignment between intention and interaction for behavior prediction. We evaluate our method on two downstream tasks, the pedestrian trajectory prediction and pedestrian intention estimation, and show that it outperforms all the existing methods.

Large-scale vision-language models (e.g., CLIP) have recently achieved remarkable performance in retrieval tasks, yet their potential for Video-based Visible-Infrared Person Re-Identification (VVI-ReID) remains largely unexplored. The primary challenges are narrowing the modality gap and leveraging spatiotemporal information in video sequences. To address the above issues, in this paper, we propose a novel cross-modality feature learning framework named X-ReID for VVI-ReID. Specifically, we first propose a Cross-modality Prototype Collaboration (CPC) to align and integrate features from different modalities, guiding the network to reduce the modality discrepancy. Then, a Multi-granularity Information Interaction (MII) is designed, incorporating short-term interactions from adjacent frames, long-term cross-frame information fusion, and cross-modality feature alignment to enhance temporal modeling and further reduce modality gaps. Finally, by integrating multi-granularity information, a robust sequence-level representation is achieved. Extensive experiments on two large-scale VVI-ReID benchmarks (i.e., HITSZ-VCM and BUPTCampus) demonstrate the superiority of our method over state-of-the-art methods.

News image captioning aims to produce journalistically informative descriptions by combining visual content with contextual cues from associated articles. Despite recent advances, existing methods struggle with three key challenges: (1) incomplete information coverage, (2) weak cross-modal alignment, and (3) suboptimal visual-entity grounding. To address these issues, we introduce MERGE, the first Multimodal Entity-aware Retrieval-augmented GEneration framework for news image captioning. MERGE constructs an entity-centric multimodal knowledge base (EMKB) that integrates textual, visual, and structured knowledge, enabling enriched background retrieval. It improves cross-modal alignment through a multistage hypothesis-caption strategy and enhances visual-entity matching via dynamic retrieval guided by image content. Extensive experiments on GoodNews and NYTimes800k show that MERGE significantly outperforms state-of-the-art baselines, with CIDEr gains of +6.84 and +1.16 in caption quality, and F1-score improvements of +4.14 and +2.64 in named entity recognition. Notably, MERGE also generalizes well to the unseen Visual News dataset, achieving +20.17 in CIDEr and +6.22 in F1-score, demonstrating strong robustness and domain adaptability.

Text-to-image person re-identification (TIReID) aims to retrieve the most relevant pedestrian images from an image gallery based on natural language descriptions. Recent studies have achieved significant performance improvements by leveraging Masked Language Modeling (MLM) to align fine-grained information through local matching. However, in the text feature extraction, randomly masking text tokens may disrupt the semantic relationships between these local tokens, leading to feature misalignment; on the other hand, from an image feature perspective, redundant patches in pedestrian images hinder the information interaction across modalities. Moreover, the presence of noisy image-text pairs further complicates the learning process, as the model may be misled into recognizing incorrect patterns. To address these issues, we propose a robust fine-grained local alignment framework based on Key Phrase Dynamic Mask (KPDM). First, we strengthen the semantic relationships between text tokens by implementing a "adjective + noun'' phrase-level masking strategy, and design a frequency-based masked language loss (FMLM) to supervise fine-grained semantic-level local alignment. Second, we integrate cross-layer importance estimation to highlight key pedestrian image representations while removing redundant image features. Third, we propose a trusted consensus partitioning mechanism, utilizing intra-identity image-text similarity distributions to identify noisy pairs, enhancing the model robustness. Extensive experiments show that our method achieves 67.95% Rank-1 and 51.88% mAP on the RSTPReid dataset, exceeding the previous state-of-the-art by 2.6% and 1%. Furthermore, KPDM achieves Rank-1 accuracies of 75.97% on the CUHK-PEDES dataset and 67.78% on the ICFG-PEDES dataset, outperforming earlier methods.

Existing unsupervised image alignment methods exhibit limited accuracy and high computational complexity. To address these challenges, we propose a dense cross-scale image alignment model. It takes into account the correlations between cross-scale features to decrease the alignment difficulty. Our model supports flexible trade-offs between accuracy and efficiency by adjusting the number of scales utilized. Additionally, we introduce a fully spatial correlation module to further improve accuracy while maintaining low computational costs. We incorporate the just noticeable difference to encourage our model to focus on image regions more sensitive to distortions, eliminating noticeable alignment errors. Extensive quantitative and qualitative experiments demonstrate that our method surpasses state-of-the-art approaches.

Scenes with water surfaces present a significant challenge for Gaussian Splatting due to the simultaneous presence of refraction and reflection, as well as the difficulty of accurately estimating the geometry of transparent water surfaces. To address this, we propose a novel framework for reconstructing scenes involving both reflection and refraction caused by water surfaces. The water surface is modeled as a trainable plane, and 2D Gaussian ray tracing is applied to account for refraction through the water. We extend 2D Gaussian Splatting by introducing a soft mask parameter and a dual set of Gaussian primitives, which handle both reflected and refracted effects. Our method achieves state-of-the-art performance on newly constructed water surface datasets, including both synthetic and real scenes, and significantly outperforms prior approaches in water-interacting regions. Furthermore, we demonstrate the editability of our model by manipulating the index of refraction to suppress or modify refractive effects, enabling scene transformations into different liquids.

Once trained, neural networks memorize information in diffusely encoded parameters, making it difficult to forget in support of the right to be forgotten. Unlearning aims to remove the influence of data, with performance measured against a retrained model that excludes the data. However, understanding the behavior of gold-standard retraining remains underexplored. We compare original and retrained models and observe that most prediction changes occur in peripheral samples near decision boundaries. Consequently, we propose PeriUn, a selective strategy that unlearns only peripheral samples to mimic retrained model behavior with minimal disruption, unlike prior works that remove the entire request. Combined with the Random Label based method, PeriUn significantly improves both generalization and privacy metrics. Specifically, on TinyImageNet with VGG16, PeriUn increases the Tug-of-War score by 22 points compared to the strongest. Besides, the MIA gap score surpasses the state-of-the-art method, improving by 8.7 points after applying PeriUn. Further analyses confirm that PeriUn better preserves the feature space and aligns closely with the retrained model.

Despite the progress made through deep learning, existing Visual Object Tracking (VOT) frameworks struggle with real-world challenges. Recent approaches incorporate additional modalities like Depth, Thermal Infrared, and Language to enhance the robustness of VOT, particularly with the improvement of the depth sensor precision, facilitating RGB-D tracking. However, current RGB-D trackers often copy RGB tracking paradigms, leading to inefficiency due to two-stream architectures that fail to exploit heterogeneous features, and reliance on simplistic or large-parameter fusion methods. To address these challenges, we propose AMTrack, a one-stream RGB-D tracker leveraging Mamba's linear complexity for simultaneous feature extraction and two-stage cross-modal feature fusion. Our innovation also includes a low-parameter Multimodal Mix Mamba (3M) module, which optimizes deep feature fusion and reduces computational overhead. The advantage of the 3M module stems from our Multimodal State Space Model (MSSM), a multimodal feature interaction component reconstructed based on SSM. Experiments across multiple RGB-D tracking datasets indicate that AMTrack achieves superior performance with lower parameters and memory demands compared to state-of-the-arts.

Introducing high-quality references can largely alleviate the uncertainty in blind face image restoration tasks, yet the equivocal utilization of reference priors makes it still a struggle to well preserve the human identity. We attribute the identity inconsistency to two deficiencies of existing reference-based face restoration methods, namely the inability to effectively determine which features need to be transferred, and the failure to preserve the structure and details of the selected features. This work mainly focuses on these two issues, and we present a novel blind face image restoration method that considers reference selection, transfer, and reconstruction (RefSTAR) to introduce proper features from reference images. Specifically, we construct a reference selection (RefSel) module, which can generate accurate masks to select reference features. For training the RefSel module, we construct a RefSel-HQ dataset through a mask generation pipeline, which contains annotated masks for 10,000 ground truth-reference pairs. To guarantee the exact introduction of selected reference features, a feature fusion paradigm is designed for reference feature transferring, and a Mask-Compatible Cycle-Consistency Loss is redesigned based on reference reconstruction to further ensure the presence of selected reference image features in the output image. Experiments on various backbone models demonstrate superior performance, showing better identity preservation ability and reference feature transfer quality.

Comprehensively and flexibly capturing the complex spatio-temporal dependencies of human motion is critical for multi-person motion prediction. Existing methods grapple with two primary limitations: i) Inflexible spatiotemporal representation due to reliance on positional encodings for capturing spatiotemporal information. ii) High computational costs stemming from the quadratic time complexity of conventional attention mechanisms. To overcome these limitations, we propose the Spatiotemporal-Untrammelled Mixture of Experts (ST-MoE), which flexibly explores complex spatio-temporal dependencies in human motion and significantly reduces computational cost. To adaptively mine complex spatio-temporal patterns from human motion, our model incorporates four distinct types of spatiotemporal experts, each specializing in capturing different spatial or temporal dependencies. To reduce the potential computational overhead while integrating multiple experts, we introduce bidirectional spatiotemporal Mamba as experts, each sharing bidirectional temporal and spatial Mamba in distinct combinations to achieve model efficiency and parameter economy. Extensive experiments on four multi-person benchmark datasets demonstrate that our approach not only outperforms state-of-art in accuracy but also reduces model parameter by 41.38% and achieves a 3.6× speedup in training.